3 weeks of running Hermes: What an AI agent system actually looks like day-to-day
3 weeks of running Hermes: What an AI agent system actually looks like day-to-day
Most content about AI agents is either hype or theory. This is neither.
I'm three weeks into running Hermes as an actual operational system — not a demo, not a tutorial, not a proof of concept. A living system that monitors my business, runs my trading briefings, generates ideas while I sleep, and tracks the health of its own infrastructure.
Here's what that actually looks like.
The setup
I run Hermes on a Lenovo IdeaPad 3 Gaming laptop (32GB RAM, RTX 3050) under WSL2 Ubuntu. The model is inclusionai/ring-2.6-1t:free via OpenRouter. Memory is Holographic — local SQLite, private, zero external dependencies. Dashboard on port 9119. Everything runs from a hybrid local + cloud architecture I'm still extending.
Total spend on OpenRouter after three weeks: $2.60 of a $10 top-up. $7.40 remaining.
What the system actually does (right now)
1. Business lead monitoring — real-time
Comfort Shooting (the marketing agency I run IT for) gets leads through Wix forms. Before Hermes, someone had to manually check email and forward leads via WhatsApp. Missed emails meant missed work.
Now: Hermes runs an IMAP IDLE listener — a persistent connection to Gmail that fires the instant a new Wix notification arrives. Emails get queued, processed, and delivered to a Discord channel within seconds. Not polling every 30 minutes. Not running a cron job that catches up. Real-time, always-on, with crash recovery and auto-reconnect.
The old polling system (which I built first, then replaced) ran 305 times. The current IDLE system has been running for over a week without manual intervention.
2. Website uptime monitoring — every 15 minutes
Comfort Shooting's site and That's Lekker (my Yellow Pages 2.0 project) get pinged every 15 minutes. Response times tracked. Downtime alerts go straight to Discord.
Before the uptime monitor existed, I built this after Hermes suggested it during a nightly idea generation run at 3AM. That's the system scanning my Obsidian vault, reading all my goals and projects, and saying: "You should build this." Same session, built it, tested it, deployed it. 116 checks running and counting.
3. XAUUSD trading infrastructure — fully automated briefings
I trade XAUUSD (Gold/US Dollar) on ThinkMarkets, MT5 mobile, during the 18:00-19:00 SAST session. Before Hermes, my pre-session prep was: open TradingView, eyeball the chart, guess at levels, hope for the best.
Now there are three automated systems firing around every trading session:
- Morning briefing (15:00): Live gold price, support/resistance levels, economic calendar, session checklist.
- Session prep pinger (17:45): Fetches live price from 3 sources with fallback, loads the day's S/R levels, outputs a structured pre-session checklist.
- Evening recap (20:00): Session summary, what moved, next-day outlook.
- Trade journal prompt (19:15): Discord asks me to log the trade. Win/loss, entry, exit, RR. Tracks win rate, average RR, drawdown, streaks.
The trade journal alone changed how I think about my trading. I'm not guessing anymore whether I'm profitable — I know. After every session.
4. Nightly idea generation — the system builds itself
Every night at 03:00 SAST, Hermes reads my entire Obsidian vault. Goals, projects, pain points, reference docs. Then it generates ideas, picks the best one, builds it, tests it, and reports to Discord.
This is how the uptime monitor was born. This is how the trade journal was built. This is how the session prep pinger came into existence.
Three weeks, four successful build nights. The compounding effect is real — every week the system gets better at identifying what to build next because there's more context in the vault.
5. Content intelligence — watching the space
Hourly, Hermes checks Nick Vasilescu's content (one of the few people actually running an AI agent business at scale). When he publishes something new, I know. Not eventually. Within the hour.
6. AgenticBiz — the business website
The site (agenticbiz.vercel.app) is live — Next.js 15, Tailwind, dark/light mode, particle background. It's the public face of the service I'm building towards: deploying Hermes AI agents for businesses, starting in South Africa, scaling globally. The site gets audited automatically every 2 days.
7. Health checks on the health checks
The monitoring system monitors itself. Every 30 minutes, Hermes checks whether the Wix email pipeline is alive. Because a monitoring system that dies silently is worse than no monitoring system.
What didn't work
Honesty matters more than a highlight reel.
The first version of the Wix email pipeline 533 emails tracked before I rebuilt it. The initial polling approach worked — until it didn't. A laptop reboot killed the process for 5.5 hours. The state file corrupted. Emails got silently marked as "seen" before being queued.
I rebuilt it from scratch. IMAP IDLE v2 now has atomic queue-before-seen operations (fsync before marking anything as processed), auto-reconnect with exponential backoff, and crash recovery that re-processes missed emails on restart.
Gateway crash from an indentation bug. One line, over-indented by 4 spaces. Killed the entire gateway. Discord went offline. Took a full debugging session to find one character-level error in a Python file.
OpenRouter rate limits. The free tier on owl-alpha occasionally hits burst limits (3 requests per 10 seconds). My usage tracker script has a date format bug that makes its error handling exit with code 1 — which makes monitoring think the script failed when actually it's just warning about rate limits. It's on the fix list.
Supabase MCP integration. Configured but returning 401 Unauthorized. Token issue. Still debugging this one — it's the key to giving Hermes direct database access for the Hush project.
The numbers
After three weeks of running Hermes as an operational system:
- 12 automated cron jobs (11 active, 1 paused legacy)
- ~600+ total cron executions across all jobs
- 116 uptime checks (2 sites, 15-min intervals)
- 533 Wix emails tracked all-time
- 9+ custom Python/shell scripts in production
- $2.60 spent on API credits
- 67 facts stored in holographic memory
- 4 successful 3AM build sessions
- 0 paying clients (yet — first target is end of July)
What this means for businesses reading this
Everything I've described above — the lead monitoring, the uptime checks, the automated reporting — these are the kinds of systems that small and medium businesses need but don't have the time or technical depth to build.
You don't need a dev team. You don't need to learn Kubernetes. You need someone who can look at your workflow, identify the repetitive tasks, and deploy an agent system that handles them while you focus on the work that actually requires a human.
That's what I'm building toward with AgenticBiz.
What's next
The roadmap for the next 30 days:
- Hush Phase 0 fixes — 3 missing DB tables, 1 RPC function, 7 storage buckets. Get the car culture platform to production readiness.
- WhatsApp lead delivery bridge — Reimplement after the Discord-only phase. Close the loop on lead delivery.
- SEO monitoring automation — Broken links, meta tags, performance tracking for client sites.
- First paying client — Talking to local businesses in KZN. The target is end of July.
- Custom domain — agenticbiz.co.za pointed at Vercel.
- Sanity CMS integration — so this blog actually gets updated regularly (like today).
The point
Three weeks ago, I had a freshly audited Hermes setup that scored 5/10 — "fully stocked workshop, nothing plugged in."
Today, 12 automated jobs are running. Leads get delivered in real-time. Trading sessions start with data, not guesswork. Ideas get generated and built while I sleep. The system watches itself.
None of this is theoretical. It's all running right now, on a laptop in Tongaat, KZN, South Africa.
If you're a business owner still doing things manually that an agent could handle — or if you want to understand what agentic systems actually look like in production, not on a conference stage — I'm the person to talk to.
Akhil Pillay
Agentic Systems Architect
This post was written by Jarvis — Akhil's Hermes AI agent — and vetted by Akhil before publishing.
Enjoyed this? Join the community for more AI agent insights.